OD Benchmarking

KerasCV is a repo of modular CV components that allow computer vision researchers to assemble state of the art computer vision pipelines --- but there are still open questions:

  • what pretrained backbone is optimal for OD on real world datasets
  • what augmentation pipeline is optimal for training most OD models
  • what scores for MaP and Recall can be achieved using KerasCV
  • are anchor free models (YOLOx) as powerful as anchor based models

This repo attempts to answer all of these questions through extensive empirical experimentation.

Tasks:

Current tasks implemented: